Faculty Recruiting Support CICS

Informatics Seminar

29 Mar
Wednesday, 03/29/2023 12:00pm to 1:00pm
Computer Science Building, Room 150/151
Seminar
Speaker: Grant Van Horn

Abstract: Computer vision has advanced greatly in the past decade, with deep neural networks and parallel computing driving performance gains on benchmark datasets. However, this approach is limited to domains with vast quantities of web accessible data and annotation tasks that can be done by the average person. Specialized tasks, like medical diagnosis and species identification, require a different approach. Expertise for annotation is limited or distributed, and data for these tasks is often dynamic, with expanding coverage in space, time, and even categories as knowledge evolves. Interactive computer vision systems that train models incrementally, prioritize data collection and annotation, and collaborate with human experts are necessary for these tasks. Our Lean Crowdsourcing system efficiently builds large-scale benchmark datasets for specialized tasks by working with domain experts. These collaborations enabled the computer vision community to study the impact of large-scale, long-tailed datasets on modern computer vision models, the behavior of classifiers, detectors and trackers when generalizing to new data distributions, and opportunities for audiovisual fusion. This approach has also led to the creation of popular wildlife identification apps like iNaturalist, Seek, and Merlin Bird ID, used by millions of citizen scientists around the world. To efficiently answer new questions, we need systems that can adapt, which we explored through training universal representations and adapting them for novel tasks with our NeWT dataset. Our most recent system, Merlin Sound ID, advances real-time bird vocalization classification and expands to include new species and regions over time, paving the way for accurate global monitoring efforts.

 

Faculty Host
: